Interview "Teaching AI Psychological Skills for Better Diagnosis and Therapies"

Dr. Eric Schulz is the Director of the Institute of Human-Centered AI at Helmholtz Munich. The AI-expert uses foundation models to understand AI-behavior through a psychological lens. His aim: advancing technology for better diagnosis and treatments - and also fullfilling his passion of getting profound insights into human cognition.

ES: I have been the director of the Institute for Human-Centered AI and a part of the Computational Health Center since December 2023. I am inspired by the interaction between the fields of cognitive science and machine learning. I want to understand and create intelligent agents such that we can harness their abilities to advance our knowledge in diverse fields such as in the medical and social sciences.

ES: We use tools from psychology, like experimental designs and computational modeling, to understand aspects of foundation models, such as their exploration or collaboration behaviors. This approach is important because these models are essentially a black box, i.e. we cannot understand them just by looking at their weights. 

ES: I see both similarities and differences between humans and machines. Machines and humans can exhibit similar behavioral patterns and learning processes, such as "Aha moments" and cognitive biases like overconfidence. However, significant differences exist, particularly in understanding and empathy. Machines lack a "Theory of Mind," making it difficult for them to comprehend and attribute mental states to themselves and others. Additionally, AI systems generate responses often without forward planning, unlike humans who plan actions with an end goal in mind.

I leverage these insights in my research by using psychological methods designed for humans to study AI behavior. This approach helps us understand AI limitations and strengths, leading to improved AI design and training. Additionally, using AI to simulate human cognitive processes on a large scale provides valuable insights into human cognition. We are currently working on a foundation model of human cognition that should respond to any cognitive task  indistinguishably from human subjects. 
 

ES: The most surprising finding in our research has been that language models can develop an understanding of complex concepts such as the laws of physics without any direct experiential learning or physical interaction. For instance, in experiments where we showed GPT-4V images of block towers and asked if they would remain standing or collapse, the model's predictions were almost as accurate as those of human participants. This suggests that AI can grasp abstract principles purely through training on large data, challenging the traditional belief that physical interaction and strong inductive biases are necessary.

ES: Foundation models have the potential to revolutionize human health by providing advanced tools for diagnosis, treatment, and personalized medicine. Specifically, we aim to create a foundation model for computational psychiatry that can analyze vast amounts of data from patient histories, genetic information, and behavioral patterns to identify behavioral and computational signature of mental health issues and predict individual responses to different treatments. By leveraging natural language processing and machine learning, these models can offer tailored therapeutic recommendations and real-time support, enhancing the accuracy and effectiveness of psychiatric care. This approach might not only improve patient outcomes but also helps clinicians make more informed decisions.

ES: I decided to join Helmholtz Munich because of its friendly and excellent research environment and the opportunity to work on things that actually matter. I have really liked the atmosphere here right from the beginning and am very happy for the frequent interactions with leading experts, whose work in computational health and medicine I find truly inspiring. Helmholtz is also building up some serious AI expertise and GPU compute, so it is simply one of the best places to join if one wants to do computational work. Additionally, I’ve always had a soft spot for Munich—where else can you find cutting-edge science, lakes, mountains, art galleries and the best beer in the world?

ES: The biggest challenge in my research is convincing people that machine psychology is a serious and valuable field. Many remain skeptical about the applicability of psychological methods to AI, questioning whether machines can truly emulate human-like thought processes. Despite this, it's worth it every day because understanding AI behavior through a psychological lens not only advances our technology but also offers profound insights into human cognition. This dual benefit makes the effort incredibly rewarding.

ES: As a student assistant, I used to work in Gerd Gigerenzer’s Max Planck Institute in Berlin. Among many other things, I was really impressed by the friendly, interactive, and inclusive environment that Gerd and his colleagues had created. To this day, I still try to have a group as vibrant, friendly, and diverse as Gerd’s group back then.

ES: A little secret about myself is that I originally wanted to be a writer. When I was 17, I even had a public reading of poems I had written, which were all quite pretentious. While my career path eventually led me to psychology and AI research, my passion for storytelling and understanding the human experience continues to inspire my work.

Latest update: May 2024